Agent-to-Agent (A2A)
LAB-03 · Implemented foundation · read-only beta · Last reviewed 29 August 2026.
A governed interoperability pattern for external agents to discover AXIO capabilities, request bounded work and receive verifiable results without blanket authority.
The question
What must sit between agent discovery and a real business action so that interoperability does not become uncontrolled delegation?
AXIO’s approach
A2A is a policy and evidence problem before it is a transport problem. Discovery may be public and read only; transactions require identity, capability authorization, human gates and verifiable receipts.
Highlights
- A public, read only MCP and A2A foundation can expose company information and service discovery without execution authority.
- The intended A2A layer separates identity, delegated authorization and policy for each capability.
- Every consequential request is treated as a proposal until an authorized human or policy gate approves it.
- The current AXIO surface is not represented as a complete deployed A2A transaction system; commercial and operational tasks remain gated and pending.
How it works
- Discover — An agent reads a public capability description, supported protocol version and input contract. Control: Discovery exposes no secrets and grants no execution permission.
- Identify — The calling principal, user delegation and target organization are resolved independently. Control: Agent identity is not assumed to equal the authority of the end user.
- Authorize — A capability policy evaluates tool, scope, data boundary, consequence and expiry. Control: Authorization applies to each capability and request, not a permanent all tools token.
- Propose — The receiving agent returns the intended action, required inputs and expected evidence. Control: Operations with high impact remain proposals until approved.
- Approve — A human or explicit business rule signs the exact action and bounded parameters. Control: Approval cannot be reused for a different object, scope or consequence.
- Execute — The selected AXIO capability runs through its normal tool, sandbox and verification policies. Control: Trust at the transport layer never bypasses safeguards at the application layer.
- Receipt — The caller receives structured status, evidence and a correlation identifier. Control: A success claim must be supported by the target system or reopened artifact.
What AXIO delivers
AXIO maps how external agents can discover and request bounded capabilities without bypassing the organization’s own authority model.
- Assessment of trust boundaries between agents and principals.
- Capability manifest, authorization and approval blueprint.
- Read only discovery prototype with a roadmap to verified transactions.
Evidence & maturity
- Implemented foundation — Machine interface for read only access: The website code includes a minimal MCP surface for public company information and service listing.
- Implemented pattern — Internal bounded tools: Cortex already uses tool registries, consequence boundaries, explicit approval and verification evidence.
- Implemented foundation — Public Agent Card and governed A2A gateway: The public Agent Card now points to a deployed Cloudflare gateway with persistent A2A task state, read-only service discovery, status lookup and cancellation. Commercial and negotiation requests remain human-gated.
- Not yet evidenced as deployed — Full A2A transactions: No OAuth delegation flow, durable task lifecycle or external transactional endpoint is claimed in this note.
- Research direction — Standards alignment: Protocol and identity choices remain subject to implementation, threat modeling and interoperability testing.
Safeguards
- Public discovery remains read only and stateless.
- External agent identity is separated from user and organizational authority.
- Scopes for each capability replace blanket access.
- Consequential actions require an explicit human or maintained policy gate.
- Every completion returns structured evidence rather than an assertion from the model alone.
Frequently asked questions
What must sit between agent discovery and a real business action so that interoperability does not become uncontrolled delegation?
A2A is a policy and evidence problem before it is a transport problem. Discovery may be public and read only; transactions require identity, capability authorization, human gates and verifiable receipts.
What does AXIO deliver in this engagement?
AXIO maps how external agents can discover and request bounded capabilities without bypassing the organization’s own authority model. Typical deliverables: Assessment of trust boundaries between agents and principals; Capability manifest, authorization and approval blueprint; Read only discovery prototype with a roadmap to verified transactions.
What is the current maturity of this work?
Implemented foundation · read-only beta. Implemented foundation: The website code includes a minimal MCP surface for public company information and service listing.
Contact & resources
Founder & CEO: Michael Vega. Email: [email protected]. Phone: +506 6300 5688. LinkedIn: linkedin.com/company/axiostaging. GitHub: github.com/VegaBuildsAI. Machine-readable resources for agents and developers: llms.txt, OpenAPI specification, MCP manifest, and agent instructions.